With the emergence of large-scale AI models both domestically and internationally, the complexity of training data has exponentially increased. In the era of trillion-parameter models, a single resource pool is no longer sufficient to accommodate the petabyte-scale data storage required for large model training scenarios. For enterprises, it is imperative to deploy distributed storage composed of multiple resource pools to meet these demands.
To address the data storage demands of large AI model training, Tianyi Cloud has introduced the Concurrent File Service HPFS (CT-HPFS, High Performance File Storage), aimed at providing a high-performance storage foundation for the AI era. This platform assists enterprises in building more efficient large-scale model training platforms based on cloud resources, enabling continuous training of large models. Tianyi Cloud's HPFS achieves concurrent data reading through distributed storage, providing up to millions of IOPS and hundreds of gigabits per second (GBPS) throughput, significantly improving data reading speed. This, in turn, greatly enhances GPU utilization and empowers the efficiency of training scenarios such as HPC and AI.
Tianyi Cloud's HPFS provides a storage solution for AI data training with high concurrency, low latency, and high throughput:
High Concurrency: Tianyi Cloud's HPFS stores data across multiple nodes while ensuring file consistency. Through distributed locking mechanisms, it enables multiple clients to simultaneously read and write different parts of the same file, significantly improving concurrency performance. Single-client single-stream IO performance can be improved by over 200%, and performance for concurrent access to the same file can be increased by over 300%.
Low latency, Tianyi Cloud HPFS leverages RDMA (Remote Direct Memory Access) technology while supporting data interaction via InfiniBand, RoCE, and TCP/IP. This significantly reduces network transmission latency, enhances data transfer efficiency, and achieves high-speed interconnection.
High throughput, Tianyi Cloud HPFS is built upon NVMe storage architecture with deep optimization, fully unleashing SSD performance and significantly improving throughput and other performance metrics. Leveraging concurrent transmission channels for parallel access enhances server capability in handling high-concurrency requests. Importantly, as the node scale increases, the system's performance scales linearly.
With its multifaceted capabilities, Tianyi Cloud HPFS serves as a high-performance parallel file storage foundation for scenarios such as autonomous driving, intelligent customer service, industrial manufacturing, gene sequencing, voice navigation, and AI training.
With its outstanding technical prowess, Tianyi Cloud HPFS clinched the first prize in the "Huacai Cup" Computing Power Application Innovation Competition, showcasing the industry's full recognition of Tianyi Cloud HPFS's innovative capabilities.
As a leading player in the cloud services arena, Tianyi Cloud remains committed to technological innovation. Our focus lies in meeting the storage needs of high-performance computing scenarios. We are dedicated to continually enhancing the performance and reliability of our products, aiming to provide customers with a robust foundation for high-performance storage. Through these efforts, we contribute to the ongoing advancement of the big model industry.
